6 SCIENTIFIC HIGHLIGHT OF THE MONTH
Many-Body van der Waals Interactionsin Biology, Chemistry, and Physics
Robert A. DiStasio Jr.1, Vivekanand V. Gobre2, and Alexandre Tkatchenko2
1Department of Chemistry, Princeton University, Princeton, NJ 08544, USA2Fritz-Haber-Institut der Max-Planck-Gesellschaft, Faradayweg 4-6, 14195, Berlin, Germany
[email protected], [email protected]
Abstract
This work presents increasing evidence that many-body van der Waals (vdW) dispersion
interactions play a crucial role in the structure, stability, and function of a wide variety of sys-
tems in biology, chemistry, and physics. We start by deriving both pairwise and many-body
interatomic methods for computing the dispersion energy by considering a system of cou-
pled quantum harmonic oscillators (QHO) within the random-phase approximation (RPA).
The resulting many-body dispersion (MBD) method contains two types of energetic con-
tributions that arise from beyond-pairwise (non-additive) interactions and electrodynamic
response screening. Applications are presented that address benchmark databases of in-
termolecular interactions, the stability of extended and globular conformations of alanine
tetrapeptide, binding in the “buckyball catcher” supramolecular host–guest complex, and
cohesion in oligoacene molecular crystals. We find that the beyond-pairwise vdW interac-
tions and electrodynamic screening are shown to play a quantitative, and sometimes even
qualitative, role in describing the properties considered herein. This highlight is concluded
with a discussion of the challenges that remain in the future development of reliable (accurate
and efficient) methods for treating many-body vdW interactions in complex materials.
1 Introduction
The relevance of van der Waals (vdW) interactions in the structure, stability, and function of
molecules and materials can hardly be overemphasized [1–5]. Ubiquitous in nature, vdW interac-
tions act at distances from just a few Angstrom to several nanometers, with recent experiments
suggesting that vdW forces can even be significantly longer-ranged [6,7]. VdW interactions are
largely responsible for the formation of the gas-phase benzene dimer at low temperatures, the
stabilization required for the formation of molecular crystals, and the binding of molecules to
proteins and DNA inside living cells. In addition, vdW interactions play a central role in the
47
fields of supramolecular chemistry and nano-materials, in which non-covalent binding is essential
for structure and functionality.
In order to enable rational predictions and design of molecular and condensed-matter materi-
als, including interfaces between them, a reliable first-principles method is required that can
describe vdW interactions both accurately and efficiently. However, an accurate description of
vdW interactions is extremely challenging, since the vdW dispersion energy arises from the cor-
related motion of electrons and must be described by quantum mechanics. The rapid increase
in computational power coupled with recent advances in the development of theoretical models
for describing vdW interactions have allowed us to achieve so-called “chemical accuracy” for
binding between small organic molecules. However, the lack of accurate and efficient methods
for treating large and complex systems hinders truly quantitative predictions of the properties
and functions of technologically and biologically relevant materials.
Many encouraging approaches have been proposed in recent years for approximately including
long-range pairwise dispersion interactions in density-functional theory (DFT) [8–17]. Despite
significant progress in the field of modeling vdW interactions, many questions still remain and
further development is required before a universally applicable method emerges. For exam-
ple, pairwise interatomic vdW methods are frequently employed to describe organic molecules
adsorbed on inorganic surfaces [18–21], ignoring the relatively strong electrodynamic response
screening present within bulk materials. On the other hand, the popular non-local vdW-DF
functionals [22–24] utilize a homogeneous dielectric approximation for the polarizability, which
is not expected to be accurate for molecules. Despite this fact, interaction energies between
small organic molecules computed with such functionals turn out to be reasonably accurate.
Understanding the physical reasons as to why these different approaches “yield good results”
outside of their expected domain of applicability is important for the development of more robust
approximations.
Interatomic pairwise dispersion approaches based on the standard C6/R6 summation formula
were popularized by the DFT-D method of Grimme [10] and are now among the most widely used
methods [9,12,13] for including the dispersion energy in DFT. Despite their simplicity, these pair-
wise models provide remarkable accuracy when applied to small molecular systems, especially
when accurate dispersion coefficients (C6) are employed for atoms in molecules [25,26]. Only re-
cently have efforts been focused on going beyond the pairwise treatment of vdW contributions,
for example, the importance of the non-additive three-body interatomic Axilrod-Teller-Muto
term [27–29] was assessed, as well as the role of non-local screening in solids [30] and molecules
adsorbed on surfaces [31]. Furthermore, an efficient and accurate interatomic many-body disper-
sion (MBD) approach has been recently proposed [32], which demonstrated that a many-body
description of vdW interactions is essential for extended molecules and molecular solids, and
the influence of many-body interactions can already become significant when considering the
binding between relatively small organic molecules [32,33].
In this highlight, we present a derivation of the pairwise and many-body interatomic dispersion
energy for an arbitrary collection of isotropic polarizable dipoles from the adiabatic connec-
tion fluctuation-dissipation (ACFD) formula, which is an exact expression for the exchange-
correlation energy. We distinguish and discuss two types of interatomic many-body contri-
48
butions to the dispersion energy, which stem from beyond-pairwise non-additive interactions
and self-consistent electrodynamic response screening. By using the ACFD formula we gain a
deeper understanding of the approximations made in interatomic approaches, in particular the
DFT+MBD method [32], providing a powerful formalism for further development of accurate
and efficient methods for the calculation of the vdW dispersion energy.
Applications of the DFT+MBD method are presented for a variety of systems, including bench-
mark databases of intermolecular interactions, the stability of extended and globular conforma-
tions of alanine tetrapeptide, the binding in the “buckyball catcher” supramolecular host–guest
complex, and the cohesive energy of several oligoacene molecular crystals. For all of these cases,
the role of the beyond-pairwise non-additive vdW interactions and electrodynamic screening
captured at the DFT+MBD level of theory is critically assessed and shown to contribute in a
quantitative, and sometimes even qualitative, fashion. We conclude this highlight with a discus-
sion of the challenges that remain in the future development of accurate and efficient methods
for treating many-body vdW interactions in materials of increasing complexity.
As the modeling of vdW interactions is currently a very active field of research, it is impos-
sible to cover all of the important developments in this highlight. For more information, we
refer interested readers to the recent Ψk highlight by Dobson and Gould, which discusses sev-
eral different approaches for computing dispersion interactions [34]; to the review by Klimes
and Michaelides on dispersion methods within DFT [35]; and to the webpage for the recent
vdW@CECAM workshop that brought together many of the key players in the development
and application of vdW-inclusive first-principles methods [36].
2 Theory
The adiabatic connection fluctuation-dissipation (ACFD) theorem provides a general and exact
expression for the exchange-correlation energy [37,38], thereby allowing for the calculation of the
dispersion energy in a seamless and accurate fashion which naturally incorporates higher-order
many-body effects. In this section, we explore the use of the ACFD theoretical framework as
a basis for the understanding and future development of interatomic pairwise and many-body
dispersion methods. Beginning with a brief derivation of the ACFD correlation energy within
the random-phase approximation (RPA), we then consider the ACFD-RPA correlation energy
for a system of quantum harmonic oscillators (QHO) interacting via the dipole-dipole potential.
We derive the well-known C6/R6 interatomic pairwise summation formula from the second-order
expansion of the ACFD-RPA correlation energy for an arbitrary collection of N QHOs, each
of which is characterized by an isotropic frequency-dependent point dipole polarizability. We
then extend our model to account for spatially distributed dipole polarizabilities and derive
modified range-separated Coulomb and dipole–dipole interaction potentials that attenuate the
short-range interactions. The self-consistent screening (SCS) method is introduced which allows
us to obtain accurate screened atomic polarizabilities that are subsequently utilized as input
for the MBD method to calculate the fully screened many-body dispersion energy. Finally, the
coupling of the MBD method with standard DFT functionals (DFT+MBD method) is achieved
by employing a range-separated Coulomb potential, and allows us to treat the full range of
49
exchange and correlation effects.
2.1 The ACFD-RPA Correlation Energy Expression
For a system of nuclei and electrons, the ACFD theorem provides us with an exact expression for
the exchange-correlation energy in terms of the density-density response function χ(r, r′, iω) [37,
38], which measures the electronic response of the system at a point r due to a frequency-
dependent electric field perturbation at a point r′. Since the focus of this work is on dispersion,
which is a quantum mechanical phenomena due to the instantaneous (dynamical) correlation
between electrons, we write the ACFD formula for the correlation energy as (Hartree atomic
units are assumed throughout):
Ec = − 1
2π
∫ ∞
0dω
∫ 1
0dλTr[(χλ(r, r
′, iω)− χ0(r, r′, iω))v(r, r′)]. (1)
In this expression, χ0(r, r′, iω) is the bare or non-interacting particle response function, which
can be computed given a set of single-particle orbitals {φi} with corresponding energies {ǫi} and
occupation numbers {fi} [39, 40] as
χ0(r, r′, iω) =
∑
ij
(fi − fj)φ∗i (r)φi(r
′)φ∗j (r′)φj(r)
ǫi − ǫj + iω, (2)
and χλ(r, r′, iω) is the interacting response function at Coulomb coupling strength λ, v(r, r′) =
|r − r′|−1 is the Coulomb potential, and Tr denotes the trace operator (or six-dimensional
integration) over the spatial variables r and r′. The interacting response function, χλ, is defined
self-consistently via the Dyson-like screening equation, χλ = χ0+χ0(λv+fxcλ )χλ, which contains
fxcλ (r, r′, iω), the exchange-correlation kernel, which must be approximated in practice.
Within the ACFD formalism, the adiabatic connection between a reference non-interacting sys-
tem (defined at λ = 0) and the fully interacting system (with λ = 1), yields the correlation
energy of the system of interest, which contains the many-body dispersion energy as well as
other electron correlation effects. This is most easily facilitated by neglecting the explicit de-
pendence of fxcλ on the coupling constant, which allows for analytic integration over λ in the
ACFD correlation energy expression in Eq. (1), and forms the basis for the most widely em-
ployed approximation for fxcλ , namely the random-phase approximation (RPA) [41,42]. In what
follows, we utilize the RPA, wherein fxcλ = 0, which has been shown to yield reliable results for
a wide variety of molecules and extended systems [43–58]. In the RPA, the ACFD correlation
energy expression can be written as a power series expansion in χ0v, following elimination of χλ
using the Dyson equation and analytical integration over λ (c.f. Eq. (1)) [59,60]:
Ec,RPA = − 1
2π
∫ ∞
0dω
∞∑
n=2
1
nTr[(χ0(r, r
′, iω)v(r, r′))n]. (3)
For a more detailed review of the RPA approach for computing the correlation energy, see
Refs. [34, 58,61] and references therein.
50
2.2 Derivation of the Long-Range Interatomic Pairwise Dispersion Energy
We now apply the ACFD-RPA approach to compute the correlation energy for a collection of
interacting QHOs representing the atoms in a molecular system of interest. In doing so, we will
first derive the standard C6/R6 interatomic pairwise summation formula for a system of two
QHOs as the second-order expansion of the ACFD-RPA correlation energy within the dipole
approximation. We will then demonstrate the validity of this formula for an arbitrary collection
of N QHOs, providing a rigorous quantum mechanical derivation of the long-range interatomic
pairwise summation formula for the dispersion energy.
In what follows, each atom p in a molecular system of interest will be mapped onto a single QHO
characterized by a position vector Rp = {xp, yp, zp} and a corresponding frequency-dependent
dipole polarizability,
αp(iω) =α0p
1 + (ω/ωp)2, (4)
which is completely determined by an isotropic static dipole polarizability, α0p ≡ αp(0), and
an effective (characteristic) excitation frequency, ωp. To evaluate the ACFD-RPA correlation
energy expression in Eq. (3), we first need the bare or non-interacting response function for
the collection of QHOs, which is assembled as a direct sum over the individual QHO response
functions, χp0(r, r
′, iω), which take on the following matrix form for a QHO located at Rp and
characterized by an isotropic point dipole polarizability [61]:
χp0(r, r
′, iω) = −αp(iω)∇rδ3(r−Rp)⊗∇r′δ
3(r′ −Rp), (5)
where δ3(r−r′) is the three-dimensional Dirac delta function, and ⊗ is the tensor (outer) product.
For the moment, we assume that the QHOs are separated by a sufficiently large distance, allowing
us to use the bare dipole-dipole interaction potential to describe the interoscillator couplings,
a condition that will be relaxed when the general case is considered in the next section. This
dipole-dipole interaction potential between oscillators p and q is straightforwardly obtained from
the bare Coulomb potential, vpq = |Rp −Rq|−1, via
Tpq =
∇Rp
⊗∇Rqvpq if p 6= q
0 if p = q(6)
and is therefore a 3× 3 second-rank tensor with components given by
Tabpq = −
3RaRb −R2pqδab
R5pq
, (7)
in which a and b represent the coordinates {x, y, z} in the Cartesian reference frame, Ra and
Rb are the respective components of the interoscillator distance Rpq, and δab is the standard
Kronecker delta function. With the individual QHO response functions and the dipole-dipole
interaction tensor as defined above, we now consider the quantity χ0v in the ACFD-RPA cor-
relation energy expression in Eq. (3), which can be represented in matrix form as the product
AT . Here, A is a diagonal 3N × 3N matrix with −αp(iω) values on the 3× 3 diagonal atomic
subblocks, representing the bare or non-interacting response function for the collection of N
QHOs. The dipole-dipole interaction matrix T is a 3N × 3N matrix comprised of the 3 × 3
blocks of the Tpq tensor defined in Eqs. (6) and (7).
51
For a system composed of two QHOs separated by a distance R = |Rp −Rq| along the z-axis
and characterized by the isotropic point dipole polarizabilities αp(iω) and αq(iω), the AT matrix
takes on the following form:
χ0v ⇔ AT =
0 0 0 −αp(iω)R3 0 0
0 0 0 0 −αp(iω)R3 0
0 0 0 0 02αp(iω)
R3
−αq(iω)R3 0 0 0 0 0
0 −αq(iω)R3 0 0 0 0
0 02αq(iω)
R3 0 0 0
. (8)
With the above matrix as input, the second-order (n = 2) term of the ACFD-RPA correlation
energy expression in Eq. (3) yields
E(2)c,RPA = − 1
2π
∫ ∞
0dω αp(iω)αq(iω)Tr[(Tpq)
2] = −Cpq6
R6, (9)
where we have used the fact that Tr[(Tpq)2] = 6/R6
pq and the Casimir-Polder integral
Cpq6 =
3
π
∫ ∞
0dω αp(iω)αq(iω) (10)
to determine the Cpq6 dispersion coefficient from the corresponding pair of frequency-dependent
dipole polarizabilities. The above equation is of course the familiar expression for the long-range
dispersion interaction between two spherical atoms.
To demonstrate the validity of this formula for an arbitrary collection of N QHOs, one needs to
consider the action of the spatial trace operator in Eq. (3) on the general 3N × 3N AT matrix.
As seen above, the second-order term in the ACFD-RPA correlation energy expansion requires
the trace of the square of the AT matrix, for which the p-th diagonal element is simply the scalar
product between the corresponding p-th column and p-th row of AT . As such, the overall trace
corresponds to an accumulated sum of the diagonal elements contained in the smaller (Tpq)2
subblocks ∀ p, q, each weighted by the product αp(iω)αq(iω). Since Tr[(Tpq)2] = 6/R6
pq for any
subblock Tpq, regardless of the geometry of the oscillator assembly, the second-order expansion
of Eq. (3) reduces to
E(2)c,RPA = −1
2
∑
pq
Cpq6
R6pq
, (11)
following the repeated use of the Casimir-Polder identity in Eq. (10) to determine the set of
interoscillator dispersion coefficients. The reader will notice that this expression is nothing more
than the standard interatomic pairwise summation formula utilized by methods such as DFT-D
to compute the dispersion energy corresponding to a collection of N atoms.
Although the second-order expansion of the ACFD-RPA correlation energy in Eq. (3) yields
the familiar interatomic pairwise expression for the dispersion energy given by Eq. (11), the
former equation is more general and provides us with a powerful formalism for the further
development of highly accurate and efficient methods for computing the dispersion energy in
molecular systems of interest. For one, the ACFD-RPA correlation energy expression allows
for the explicit utilization of the tensor form of the frequency-dependent dipole polarizability,
enabling a fully anisotropic treatment of the dispersion energy. In this regard, anisotropy in
52
the polarizability has been found to play a non-negligible role in the accurate description of
intermolecular dispersion interactions [62, 63]. In the next two sections, we extend our above
treatment by considering QHOs characterized by spatially distributed (instead of point) dipole
polarizabilities and describe a method for capturing the anisotropy in the frequency-dependent
dipole polarizability based on the solution of the self-consistent Dyson-like screening equation of
classical electrodynamics. Secondly, the use of the untruncated ACFD-RPA correlation energy
expression allows for the explicit inclusion of the higher-order (n > 2) energetic contributions
that arise naturally in the power series expansion of χ0v. These terms include two distinct
energetic contributions: the beyond-pairwise (non-additive) many-body interactions (to N th
order) and the higher-order electrodynamic response screening (to infinite order). The first
example of the beyond-pairwise many-body interactions is captured in the third-order expansion
of the ACFD-RPA correlation energy (for a system with N ≥ 3), which is the so-called Axilrod-
Teller-Muto triple-dipole term [64]. The higher-order response screening is most easily illustrated
by considering a system composed of two QHOs p and q and expanding Eq. (3):
Ec,RPA = − 1
2π
∫ ∞
0dω
(6αp(iω)αq(iω)
R6pq
+9α2
p(iω)α2q(iω)
R12pq
+ . . .
), (12)
in which the second-order term corresponds to the “standard” C6/R6 pairwise dispersion inter-
action and the higher-order terms (which only survive with even powers of n) correspond to the
electrodynamic screening of the polarizability of atom p by the presence of atom q and vice versa.
In section 2.5, we again extend our above treatment and describe a method that accurately and
efficiently accounts for both beyond-pairwise non-additive many-body and higher-order electro-
dynamic response screening contributions to the dispersion energy for an arbitrary system of N
QHOs.
2.3 Extension to Spatially Distributed Dipole Polarizabilities
Correlation energy calculations carried out using the ACFD-RPA formula typically employ the
bare response function, χ0, computed using the set of occupied and virtual (unoccupied) single-
particle orbitals obtained from self-consistent Hartree-Fock, semi-local DFT, or hybrid DFT
calculations via Eq. (2). When constructed in this fashion, χ0 accounts for orbital product
(overlap) effects between the single-particle occupied and virtual states and is therefore a rel-
atively delocalized object. On the other hand, when χ0 is completely localized, its real-space
matrix representation is diagonal in form, reflecting the fact that orbital product (overlap) ef-
fects have been neglected in the bare response function—this was the case for the A matrix
corresponding to the collection of QHOs considered in the previous section. In this limit of fluc-
tuating point dipoles, the interaction between QHOs diverges when the interoscillator distances
become relatively close.
In the previous section, we assumed that the QHOs were separated by a sufficiently large distance
to allow us to describe the interactions between them using the bare dipole-dipole potential, a
condition that will now be relaxed in order to consider the general case, in which QHOs can
be separated by typical chemical bond distances. The most straightforward way to avoid the
near-field divergence is to incorporate orbital product (overlap) effects for the set of QHOs
53
through a modification of the interaction potential at short interoscillator distances. Therefore,
instead of using the bare Coulomb potential to derive the dipole-dipole interaction tensor, we
will utilize a modified Coulomb potential that (i) accounts for orbital product (overlap) effects
at short interoscillator distances and (ii) becomes equivalent to the bare Coulomb potential
in the long-range. This range-separated Coulomb potential can actually be rigorously derived
from first principles by utilizing fundamental quantum mechanics, i.e., the solutions of the
Schrodinger equation for the QHO. In order to proceed, we first note that the ground state
QHO wavefunction, ψQHO0 (r), is a spherical Gaussian function and hence the corresponding
ground state QHO charge density is also a spherical Gaussian function by the Gaussian product
theorem, i.e.,
nQHO0 (r) = |ψQHO
0 (r)|2 =exp[−r2/2σ2]
π3/2σ3, (13)
in which σ represents the width or spread of the Gaussian. The corresponding Coulomb inter-
action between two spherical Gaussian charge distributions associated with oscillators p and q
can then be derived as [65]
vpq =erf[Rpq/σpq]
Rpq, (14)
in which σpq =√σ2p + σ2q , is an effective width obtained from the Gaussian widths of oscillators
p and q, that essentially determines the correlation length of this interaction potential. Since
the dipole polarizability relates the response of a dipole moment to an applied electric field, the
σ parameters physically correspond to the spatial spread of the local dipole moment distribu-
tion centered on a given oscillator. In fact, these Gaussian widths are directly related to the
polarizability in classical electrodynamics [66] and can be derived from the dipole self-energy
(i.e., the zero-distance limit of the dipole-dipole interaction potential derived below in Eq. (15))
as σp = (√
2/π αp/3)1/3.
From Eq. (14), it is clear that this modified Coulomb potential satisfies both of the aforemen-
tioned conditions, so we now proceed to derive the dipole-dipole interaction tensor between two
QHOs p and q from this modified Coulomb potential, which takes on the following form after
straightforward algebra (c.f., Eqs. (6) and (14)):
Tabpq = ∇Rp
⊗∇Rqvpq
= −3RaRb −R2
pqδab
R5pq
(erf[Rpq/σpq]−
2√π
Rpq
σpqexp[−(Rpq/σpq)
2]
)
+4√π
RaRb
σ3pqR2pq
exp[−(Rpq/σpq)2]. (15)
We note that the above expression describes a potential that (i) attenuates the interaction be-
tween oscillators at short distances in comparison to the bare dipole-dipole interaction potential,
converging to a finite value even in the zero-distance limit, and (ii) becomes equivalent to the
bare dipole-dipole interaction potential for large interoscillator distances. Hence, the use of this
range-separated dipole-dipole interaction potential for an arbitrary collection of QHOs not only
allows us to avoid the near-field divergence that plagues the short-range, but also provides us
with the simultaneous ability to correctly describe the long-range dispersion energy.
54
2.4 Electrodynamic Response Screening and Polarizability Anisotropy
Neglecting retardation effects due to the finite speed of light, the long-range dispersion energy
between two atoms in vacuo originates from the electrodynamic interaction of “atomic” dipolar
fluctuations. However, when an atom is embedded in a condensed phase (or in a sufficiently
large molecule), the corresponding dipolar fluctuations significantly differ from the free atom
case, and in fact, this difference originates from both the local chemical environment surround-
ing the atom and the long-range electrodynamic interaction with the more distant fluctuating
dipoles decaying via a ∼ 1/R3 power law. In other words, each atom located inside a molecule
or material experiences a dynamic internal electric field created by both the local and non-local
fluctuations associated with the surrounding atoms. Depending on the underlying topology
of the chemical environment, this fluxional internal electric field can give rise to either polar-
ization or depolarization effects, and is largely responsible for the anisotropy in the molecular
polarizability tensor [67, 68]. Therefore, it is essential to include the environmental screening
effects arising from both the short- and long-range in accurate first-principles calculations of the
dispersion energy.
To address this issue, we again represent the N atoms in a given molecular system of interest
as a collection of N QHOs, each of which is characterized by an isotropic frequency-dependent
point dipole polarizability, the form of which is given in Eq. (4). Up to this point, we have yet
to specify the parameters necessary to construct this polarizability for a given QHO, namely, an
isotropic static dipole polarizability, α0p, and an effective excitation frequency, ωp. From the dis-
cussion above, it is clear that we need to incorporate both short- and long-range environmental
effects in our description of the QHO frequency-dependent polarizability in order to accurately
capture the electrodynamic response screening and anisotropy effects. To account for the lo-
cal chemical environment, we utilize the Tkatchenko-Scheffler (TS) prescription [25], in which
α0p[n(r)] and ωp[n(r)] are defined as functionals of the ground-state electron density, obtained
from an initial self-consistent quantum mechanical calculation using either semi-local or hybrid
DFT [69]—methods that can accurately treat electrostatics, induction, exchange-repulsion, and
local hybridization effects, but lack the ability to describe long-range dispersion interactions [70].
Assuming that the system (whether it be an individual molecule, a collection of molecules, or
even condensed matter), has a finite electronic gap and can therefore be divided into effective
atomic fragments, the Hirshfeld, or stockholder [71], partitioning of the electron density is then
utilized to account for the local chemical environment surrounding each atom. Since both param-
eters are referenced to highly accurate free-atom reference data, short-range quantum mechanical
exchange-correlation effects are accounted for in these quantities by construction. In fact, the
frequency-dependent polarizabilties defined in this manner yield C6 coefficients that are accurate
to 5.5% when compared to reference experimental values for an extensive database of atomic
and (small) molecular dimers. Nevertheless, this parameterization of the frequency-dependent
dipole polarizability clearly lacks the aforementioned long-range electrodynamic screening that
extends beyond the range of the exponentially decaying atomic densities, and these effects must
be accounted for self-consistently within this system of fluctuating oscillators.
To accurately capture the long-range electrodynamic response screening and anisotropy effects,
we self-consistently solve the Dyson-like screening equation utilizing the range-separated dipole-
55
dipole interaction tensor derived above in Eq. (15), thereby improving upon our initial de-
scription of the bare response function corresponding to this collection of QHOs. To proceed
forward, we recall that this initial bare response function, χ0, was constructed as a direct sum
over the individual oscillator response functions given in Eq. (5), i.e., χ0 = χp0 ⊕ χq
0 ⊕ . . ., which
now correspond to QHOs characterized by isotropic frequency-dependent dipole polarizabilities
parameterized using the TS definitions for α0p[n(r)] and ωp[n(r)] presented above. Therefore,
the real-space matrix representation of χ0 is diagonal and as a result, the corresponding self-
consistent Dyson-like screening (SCS) equation is separable and can be recast as the following
nonhomogeneous system of linear equations for a given frequency ω:
αp(iω) = αp(iω)− αp(iω)N∑
q 6=p
Tpqαq(iω) p = 1, 2, . . . , N. (16)
In Eq. (16), the complexity associated with integrating over spatial variables r and r′ has been
absorbed into Tpq, the 3 × 3 block of the range-separated dipole-dipole interaction tensor in
Eq. (15), which facilitates the use of overlapping spatially distributed frequency-dependent dipole
polarizabilities by eliminating the issues associated with the near-field divergence. The set
of αp(iω) are the unknowns in the SCS equation and physically correspond to QHO dipole
polarizabilities that account for both short-range (via the TS scheme) and long-range (via the
solution of the SCS equation) electrodynamic response screening effects arising from the chemical
environment. The solution of Eq. (16) with input polarizabilities computed at the TS level will
be referred to as TS+SCS throughout the remainder of this work.
The SCS equation can be solved in matrix form via inversion of the 3N × 3N Hermitian A
matrix, which contains the inverse of the atomic frequency-dependent dipole polarizability ten-
sors, α−1p (iω), along the diagonal 3× 3 atomic subblocks, and the range-separated dipole-dipole
coupling tensor, Tpq, in each of the corresponding 3 × 3 non-diagonal subblocks. Inversion
of A yields the fully screened polarizability matrix, B, from which one can obtain the fully
screened molecular polarizability tensor by internally contracting over all atomic subblocks,
α =∑N
pqBpq and the fully screened set of atomic polarizability tensors by partial internal con-
traction αp =∑N
q Bpq. From these screened atomic polarizability tensors, α0p and ωp can be
obtained for each atom as described in Ref. [32] and will be used to compute the full many-body
dispersion energy in the next section.
2.5 The Many-Body Dispersion Energy: The DFT+MBD Method
To compute the full many-body dispersion (MBD) energy, we represent the N atoms in a given
molecular system of interest as a collection of QHOs, each of which is now characterized by
a screened static dipole polarizability, α0p, and screened excitation frequency, ωp, computed
at the TS+SCS level of theory. We then directly solve the Schrodinger equation for this set
of fluctuating and interacting QHOs within the dipole approximation, with the corresponding
Hamiltonian [72–76]:
H = −1
2
N∑
p=1
∇2µp
+1
2
N∑
p=1
ω2pµ
2p +
N∑
p>q
ωpωq
√α0pα
0qµpTpqµq, (17)
56
in which µp =√mpξp is defined in terms of ξp, the displacement of a given QHO p from
its equilibrium position, and mp = (α0pω
2p)
−1. In Eq. (17), the first two terms correspond to
the single-particle kinetic and potential energy, respectively. The last term in the Hamiltonian
describes the coupling between QHOs via the dipole-dipole interaction tensor (Tpq = ∇Rp⊗
∇RqW (Rpq), where W (Rpq) will be defined below).
For a system of coupled QHOs, we have in fact proven the equivalence [77] between the full
interaction energy obtained from the diagonalization of the Hamiltonian in Eq. (17) and the
ACFD-RPA correlation energy expression in Eq. (3). Therefore, the full ACFD-RPA correlation
energy can be efficiently computed by diagonalizing the 3N × 3N Hamiltonian matrix. Hence,
the MBD energy is computed as the difference between the zero-point energies of the coupled
(collective) and uncoupled (single-particle) QHO frequencies, i.e.,
EMBD =1
2
3N∑
p=1
√λp −
3
2
N∑
p=1
ωp = Ec,RPA, (18)
in which λp are the Hamiltonian matrix eigenvalues.
Although the MBD energy is part of the long-range correlation energy, the full correlation energy
in general also includes other contributions. In order to construct an electronic structure method
that treats the full range of exchange and correlation effects, we need to couple the MBD energy
in Eq. (18) to an approximate semilocal DFT functional. Instead of utilizing an ad hoc damping
function, as typically employed in interatomic pairwise approaches, the coupling of MBD to
an underlying functional (DFT+MBD) is achieved via the following range-separated Coulomb
potential [78,79], which suppresses the short-range interactions that are already captured at the
DFT level,
W (Rpq) =(1− exp(−(Rpq/R
vdWpq )β)
)/Rpq, (19)
where β is a range-separation parameter that controls how quickly W (Rpq) reaches the long-
range 1/Rpq asymptote, and RvdWpq = R
vdWp + R
vdWq are the screened vdW radii as defined in
Refs. [25, 32].
The value of the single range-separation parameter, β, is obtained from global optimization of
the total DFT+MBD energy on the S22 test set, a widely employed benchmark database of
noncovalent intermolecular interactions [80,81]. For the PBE [82] and PBE0 [83,84] functionals,
the optimized values of the β parameter were found as 2.56 and 2.53, respectively.
Finally, we remark that our choice of using the screened α0p and ωp parameters as input in
Eq. (17) is not unique. Other choices are certainly possible from the viewpoint of the ACFD
formula and we will investigate such alternatives in more detail in future work. In addition, the
coupling of the long-range MBD energy to a semilocal or hybrid DFT functional distinguishes
the DFT+MBD method from the widely used RPA@DFT approaches [43–58] for computing
the electron correlation energy. Furthermore, the MBD energy can be efficiently computed by
diagonalizing the 3N × 3N Hamiltonian matrix, enabling MBD calculations for thousands of
atoms on a single processor.
57
3 Applications: The Role of Electrodynamic Response Screen-
ing
The simplest possible model for the polarizability of molecules and solids consists of a sum over
effective hybridized polarizable atoms, as given by Eq. (4). This model can be very effective in
reproducing accurately known isotropic molecular polarizabilities and isotropic C6 coefficients.
For example, the TS method uses a localized atom-based model and yields an accuracy of ≈ 14%
for the isotropic polarizabilities of more than 200 molecules [85] and 5.5% for the C6 coefficients
in 1225 cases [25]. However, one has to recognize that the polarizability measures the response of
a dipole moment to an applied electric field. Since both the dipole moment and the electric field
are vector quantities, the dipole polarizability is evidently anisotropic and should be described
by a second-rank tensor. Hence, the rather simplified additive model fails to correctly capture
the anisotropy in the molecular polarizability [2]. Within the framework of electronic structure
calculations, the static polarizability can be computed as the second derivative of the total
energy with respect to an applied electric field. An alternative, but equivalent formulation for
computing the polarizability is based on the fact that the single-particle orbitals in a molecule are
electrodynamically coupled. The solution of the coupling equations leads to the many-electron
frequency-dependent polarizability of the full system.
The TS+SCS method introduced above in Eq. (16) is based on such an electrodynamic in-
teraction model. Upon obtaining effective isotropic parameters for atoms in a molecule or a
solid from the ground-state electron density, the non-local polarizability tensor is determined
from the solution of a system of dipole–dipole coupling equations. The dipole–dipole coupling
between atoms naturally introduces anisotropy in the molecular polarizability, even if we start
with purely isotropic atomic polarizabilities. We now illustrate the importance of electrody-
namic screening for three different cases: small and medium-size molecules, a linear chain of H2
molecules, and silicon clusters of increasing size.
3.1 Small and Medium-Size Molecules
Table 1 shows the three components of the molecular static polarizability, αxx, αyy, and αzz,
along with the isotropic static polarizability, αiso, for a database of 18 molecules [68]. The TS
atomic partitioning of the polarizability integrated in different directions yields a mean absolute
error of 13.2% for the isotropic molecular polarizability, and a much larger error of 76.3% for
the fractional anisotropy (FA), defined as
FA =
√1
2
(αxx − αyy)2 + (αxx − αzz)2 + (αyy − αzz)2
α2xx + α2
yy + α2zz
. (20)
Upon including screening effects using the TS+SCS model [Eq. (16)], the isotropic polarizability
is improved to 9.1%, and, more importantly, the accuracy of FA is improved by a factor of two
to 33.5%. We suggest that a substantial part of the remaining error stems from the isotropic
input into the SCS model. Using the full electron density anisotropy at the TS level requires a
substantial extension of the TS+SCS model, which is work that is currently in progress. We note
that for the calculation of the vdW energy, what matters is the integration of the polarizability
58
over imaginary frequencies, α(iω), hence the error in the static polarizability is ameliorated when
computing the vdW energy.
Table 1: The isotropic polarizability αiso, along with its three components αxx, αyy, and αzz
(in bohr3) for a database of molecules with reference data taken from Ref. [68]. The MARE for
the components corresponds to the error in the fractional anisotropy (see text). The results are
reported for the TS method (with anisotropy computed from the Hirshfeld partitioning, where
the r3 operator is partitioned as (xx+ yy + zz)r), and the TS+SCS method.
Experiment TS TS+SCS
molecule αiso αxx αyy αzz αiso αxx αyy αzz αiso αxx αyy αzz
H2 5.33 4.86 4.86 6.28 4.61 4.57 4.63 4.63 3.98 3.15 3.15 5.64
N2 11.88 9.79 9.79 16.06 12.59 12.02 12.02 13.73 11.24 8.79 8.79 16.14
O2 10.80 8.17 8.17 15.86 10.03 10.02 10.02 10.06 9.86 7.61 7.61 14.36
CO 13.16 11.00 11.00 17.55 14.62 13.80 13.80 16.27 13.21 10.76 10.76 18.13
ethane 30.23 26.86 26.86 37.05 33.72 33.18 33.18 34.79 31.86 28.78 28.79 38.02
propane 43.05 38.74 38.74 51.69 49.04 47.68 48.88 50.55 46.66 39.75 42.79 57.43
cyclopentane 61.75 56.69 61.88 66.67 74.56 72.49 75.56 75.63 68.49 57.54 73.95 73.98
cyclohexane 74.23 63.30 79.70 79.70 90.59 88.36 91.70 91.70 83.27 67.91 90.96 90.96
dimethylether 35.36 29.63 33.34 43.05 39.24 38.53 39.47 39.70 37.82 32.11 32.70 48.66
P-dioxane 58.04 47.24 63.43 63.43 70.50 69.68 70.17 71.64 65.76 53.12 67.20 76.97
methanol 22.40 17.88 21.80 27.60 24.44 23.99 24.61 24.72 23.11 19.96 21.44 27.92
ethanol 34.28 30.37 33.61 38.87 39.71 38.73 39.15 41.23 37.64 32.33 37.28 43.29
formaldehyde 16.53 12.35 18.63 18.63 19.06 17.09 19.54 20.55 18.09 11.42 18.86 24.00
acetone 43.12 29.83 49.74 49.74 49.07 45.80 50.22 51.18 48.05 35.41 49.90 58.83
acetonitrile 30.23 25.98 25.98 38.74 32.51 31.17 31.17 35.19 32.82 23.62 23.62 51.22
(CH3)3CCN 64.72 60.94 60.94 72.27 79.13 78.16 78.16 81.07 77.09 70.65 70.65 89.98
methane 17.68 17.68 17.68 17.68 18.90 18.90 18.90 18.90 17.39 17.39 17.39 17.39
benzene 69.70 45.10 82.00 82.00 75.29 71.82 77.02 77.03 71.95 33.02 91.41 91.42
MARE - - 13.2% 76.3% 9.1% 33.5%
For a pair of atoms or molecules A and B, the CAB6 coefficient determines their long-range vdW
interaction energy. One of the main achievements of the TS method consists of a parameter-free
definition to determine the CAB6 coefficients with an accuracy of 5.5% for a broad variety of
small and medium-size molecules (1225 CAB6 coefficients). The performance of the TS method
is shown in Figure 1, where a remarkable correlation can be seen with reliable CAB6 values com-
puted from the experimental dipole-oscillator strength distributions (see Ref. [25] for a detailed
analysis). The reason behind such a good performance is that SCS effects beyond semilocal
hybridization largely average out when computing C6 coefficients for small molecules. In fact,
the TS+SCS method yields an accuracy of 6.3% for the aforementioned 1225 C6 coefficients
and its performance is also shown in Figure 1. We attribute the slight increase of the error
with respect to TS as stemming from the approximation of the dipole moment distribution
by a single isotropic QHO. The largest errors of TS+SCS are found for linear alkane chains,
where the anisotropy along the chain is overestimated. Full tensor formulation of the input TS
59
Figure 1: Isotropic C6 coefficients for a database of 50 atoms and molecules (1225 data points)
computed with TS and TS+SCS methods, compared with reliable DOSD values (see text).
polarizabilities is under way and preliminary results indicate that the molecular anisotropy is
improved.
3.2 Linear Chain of H2 Molecules
We further illustrate the importance of SCS effects with the example of the linear (H2)3 chain,
consisting of three H2 dimers with alternating bond lengths (2 bohr inside the dimer and 3
bohr between the dimers). An accurate calculation of the polarizability of such hydrogen dimer
chains is considered to be a significant challenge for electronic structure theory [86]. We have
calculated the reference frequency-dependent polarizability for (H2)3 using the linear-response
coupled-cluster method (LR-CCSD) as implemented in the NWChem code [87, 88]. The LR-
CCSD method is a state-of-the-art approach for computing static and frequency-dependent
molecular polarizabilities, and it yields results that agree to ≈3% when compared to reliable
experimental values. The results for the isotropic and anisotropic C6 coefficients for this chain
at the TS, TS+SCS, and LR-CCSD levels of theory are shown in Table 2. The TS method
yields a vanishingly small anisotropy in the C6 coefficient since it only accounts for the local
environment. On the contrary, TS+SCS correctly captures the dipole alignment (polarization)
along the (H2)3 chain, leading to a significant anisotropy that is in fair agreement with the
reference LR-CCSD values. Also, the isotropic C6 coefficient is noticeably improved when using
the TS+SCS approach in comparison to TS.
60
Table 2: Anisotropic (C6,⊥, C6,||) and isotropic (C6,iso) C6 coefficients for the linear (H2)3 chain
using the TS and TS+SCS methods. Reference linear-response coupled-cluster (LR-CCSD)
results are also shown. All values in hartree·bohr6.C6,⊥ C6,|| C6,iso
TS 166 161 165
TS+SCS 89 692 223
LR-CCSD 115 638 238
3.3 Silicon Clusters
We have shown that the TS+SCS method can rather effectively describe the anisotropy and po-
larization effects in molecules. We now illustrate that the TS+SCS approach can also accurately
treat depolarization in clusters and solids with the example of hydrogen-saturated silicon clus-
ters of increasing size. The cluster–cluster C6 coefficients are shown in Figure 2. The reference
values correspond to the TDLDA calculations of S. Botti et al. [89]. We measured the accuracy
of TDLDA using the experimentally derived C6 coefficient for the SiH4 molecule [90] and the
CSi−Si6 coefficient in the silicon bulk determined from the Clausius-Mossotti equation with the
experimental dielectric function. For the SiH4 molecule, TDLDA yields a 13% overestimation
and this error is further reduced to 3% for the silicon bulk. Therefore, we deem the TDLDA C6
coefficients as good references for the larger silicon clusters. For smaller clusters, the TS values
are accurate and are in good agreement with experiment and TDLDA as expected. However, the
error in the TS method increases progressively with the cluster size. For the largest Si172H120
cluster, the TS approach yields an overestimation of 27%. TS+SCS leads to an overall depolar-
ization for the larger clusters, decreasing the error significantly in comparison to TDLDA. The
depolarization effect is even larger for the Si bulk. The TS scheme yields an overestimation of
68% in the CSi−Si6 coefficient in comparison to the value derived from the experimental dielectric
function, while the TS+SCS approach reduces the overestimation to just 8%.
4 Applications: Performance of the DFT+MBD Method
Having established the accuracy of the TS+SCS method for computing the vdW coefficients
for a wide variety of systems from molecules to solids, we now assess the performance of the
DFT+MBD method based on the TS+SCS input (see Section 2.5) for a broad variety of molecu-
lar systems. The cases studied herein include the binding energies of molecular dimers, conforma-
tional energetics of extended and globular alanine tetrapeptide, binding in the supramolecular
host–guest buckyball catcher complex, as well as cohesion in molecular crystals composed of
oligoacenes. The all-electron numeric atom-centered orbital code FHI-aims [91] was utilized for
the DFT calculations discussed in this work.
61
Figure 2: Cluster–cluster isotropic C6 coefficients for hydrogen-terminated silicon clusters of
increasing size. The TDLDA results are from Ref. [89].
4.1 Intermolecular Interactions: The S22 and S66 Databases
In order to assess the performance of the DFT+MBD method, we first study the S22 database
of intermolecular interactions [80], a widely used benchmark database for which reliable binding
energies have been calculated using high-level quantum chemical methods [80,81]. In particular,
we use the recent basis-set extrapolated CCSD(T) binding energies calculated by Takatani et
al. [81]. These binding energies are presumed to have an accuracy of ≈ 0.1 kcal/mol (1%
relative error), and this level of accuracy is required for an unbiased assessment of approximate
approaches for treating dispersion interactions.
Figure 3 shows the performance of the DFT+MBD method on the S22 database when used
with the standard semilocal PBE [82] functional and the hybrid PBE0 [83,84] functional which
includes 25% Hartree-Fock exchange. The inclusion of the many-body vdW energy leads to a
remarkable improvement in accuracy compared to the PBE+TS-vdW method [25]. The largest
improvement when using the MBD energy over the pairwise TS-vdW energy is observed for the
methane dimer and the parallel-displaced benzene dimer. We note that the methane dimer is
bound by only 0.53 kcal/mol at the CCSD(T) level of theory, and the MBD energy reduces the
binding by 0.19 kcal/mol with respect to TS-vdW, explaining the large reduction in error seen
in Figure 3. This reduction does not come mainly from the many-body dispersion energy, rather
it is due to a more physical definition of the short-range interactions in the MBD method arising
from a range-separated Coulomb potential [32]. Taking the second-order expansion of the MBD
energy, which yields a strictly pairwise energy, leads to a change of only 0.05 kcal/mol [77]
compared to the full MBD energy. This simple test illustrates that the main difference between
PBE+TS-vdW and PBE+MBD for the methane dimer stems from the different way of treating
the short-range dispersion interactions. In addition, the inclusion of Hartree-Fock exchange in
the PBE0 functional allows for a better description of permanent electrostatic moments and
static polarizabilities for molecules, and leads to improved binding energies when compared to
the semilocal PBE functional. We note that there are two systems in the S22 database for
62
Figure 3: The performance of the PBE+TS-vdW method of Tkatchenko and Scheffler [25],
PBE+MBD, and PBE0+MBD methods on the S22 database of intermolecular interactions.
The error is reported to the basis-set converged CCSD(T) results of Takatani et al. [81].
which the relative PBE0+MBD error exceeds 10% when compared to the CCSD(T) binding
energies: pyrazine dimer (system 12) and ethene–ethyne (system 16). We attribute this finding
to the remaining inaccuracy in the anisotropy for the molecular polarizabilities computed with
the TS+SCS method. This issue will be analyzed in more detail for the case of the buckyball
catcher complex below.
To put the performance of the DFT+MBD method in the context of other currently available
approaches, we show the mean absolute relative errors (MARE) on the S22 database for a vari-
ety of state-of-the-art methods in Table 3 and in Figure 4. The number of empirical parameters
employed for the dispersion energy in every method is also enumerated in Table 3. Only the
PBE0+MBD method [32] and the rPW86+cPBE+VV10 approach [24,92] yield consistent per-
formance with errors below 6% with respect to the CCSD(T) reference data for all interaction
types. We note that the rPW86+cPBE+VV10 method uses two empirical parameters in the
expression for the dispersion energy, while the PBE0+MBD method uses only a single range-
separation parameter for the coupling of the long-range dispersion energy to the underlying DFT
functional.
Recently, Hobza’s group has significantly revised and extended the S22 database to include a
broader variety of molecules and intermolecular interactions. The result of this effort is the
so-called S66 database, composed of 66 molecular dimers [93]. The reference binding energies
for the S66 database have been computed at the CCSD(T) level of theory employing medium-
size basis sets, with an expected accuracy of ≈2-3% from the basis set limit. In order to cover
non-equilibrium geometries, CCSD(T) binding energies have also been computed for 8 different
intermolecular separations, ranging from a factor of 0.9 to 2.0 of the equilibrium distances.
Therefore, the so-called S66x8 database contains binding energies for a total of 528 complexes
63
Table 3: Performance of different methods on the S22 database of intermolecular interactions,
measured in terms of the mean absolute relative error (MARE, in %). The errors are measured
with respect to the basis-set extrapolated CCSD(T) calculations of Takatani et al. [81]. The error
is reported for hydrogen-bonded (H-B), dispersion-bonded (D-B), and mixed (M-B) systems.
The number of empirical parameters used in every approach is shown in the “N. param.” column.
Results are shown for MP2, EX+cRPA, EX+cRPA+SE [57], vdW-DF1 and vdW-DF2 [23],
rPW86+cPBE+VV10 [24,92], PBE0-D3 [28], PBE0+TS-vdW [25,26], and PBE0+MBD [32].
Method H-B D-B M-B Overall N. param.
MP2 1.8 37.4 14.8 18.9 0
EX+cRPA 11.2 21.6 14.8 16.1 0
vdW-DF1 15.2 13.0 10.8 13.0 0
PBE0-D3(Grimme) 8.4 15.5 12.7 12.3 > 3
EX+cRPA+SE 5.9 11.6 5.4 7.8 0
vdW-DF2 5.3 6.8 10.8 7.6 1
PBE0+TS-vdW 3.4 12.0 6.0 7.3 2
rPW86+cPBE+VV10 6.1 2.6 4.8 4.4 2
PBE0+MBD 4.1 3.4 5.1 4.2 1
computed at the CCSD(T) level of theory. The performance of the PBE0+MBD approach on
the S66 database is comparable to the S22 results presented above. For equilibrium geometries
in the S66 database, the mean absolute error (MAE) and MARE of the PBE0+MBD method are
0.38 kcal/mol and 6.1%, respectively. When all 528 equilibrium and non-equilibrium complexes
are taken into account, the calculated MAE and MARE are 0.37 kcal/mol and 8.5%, respectively.
The increase in the MARE stems from the S66(0.9x) and S66(0.95x) complexes with shorter-
than-equilibrium interaction distances. This is a well-known weakness of all dispersion-inclusive
DFT methods, with errors increasing when considering shorter distances, since the dispersion
energy contribution for such distances becomes very small.
We conclude that the MBD energy beyond the standard pairwise approximation is important
even when studying the binding between rather small molecules. Empirical pairwise methods
for the dispersion energy mimic some of the higher-order effects by adjusting sufficiently flexible
damping functions, but this strategy is prone to fail for different molecular conformations and
for more complex molecular geometries. We illustrate one such case in the next subsection.
4.2 Intramolecular Interactions: Conformational Energies of Alanine Tetrapep-
tide
The study of biomolecules in the gas phase corresponds to ideal “clean room” conditions, and
recent progress in experimental gas-phase spectroscopy has yielded increasingly refined vibra-
tional spectra for peptide secondary structures [94–96]. Joint experimental and ab initio theoret-
ical studies can now successfully determine the geometries of small gas-phase peptides [97–99].
Polyalanine is a particularly good model system due to its high propensity to form helical struc-
tures [100], and its widespread use as a benchmark system for peptide stability in experiments
64
Figure 4: Performance of different methods on the S22 database of intermolecular interactions,
measured in terms of the mean absolute relative error (MARE, in %). The errors are mea-
sured with respect to the basis-set extrapolated CCSD(T) calculations of Takatani et al. [81].
Results are shown for MP2, EX+cRPA, EX+cRPA+SE [57], vdW-DF1 and vdW-DF2 [23],
rPW86+cPBE+VV10 [24,92], PBE0-D3 [28], PBE0+TS-vdW [25,26], and PBE0+MBD [32].
and theory.
Here we assess the accuracy of the PBE0+MBDmethod for 27 conformations of alanine tetrapep-
tide (Ace-Ala3-NMe, for brevity called Ala4 here), for which benchmark CCSD(T) confor-
mational energies were computed in Ref. [101], based on converged MP2/CBS values from
Refs. [102, 103]. The Ala4 conformations range from a β-sheet-like fully extended structure to
a globular (“folded”) conformer. The wide variety of interactions present in peptides ranging
from hydrogen bonds to dispersion and electrostatics makes an accurate prediction of the con-
formational hierarchy of these systems quite a daunting task for affordable electronic structure
calculations. We illustrate the performance of PBE0+TS-vdW and PBE0+MBD for Ala4 con-
formers in Figure 5. The PBE0+MBD method predicts a MAE of 0.29 kcal/mol with respect to
the CCSD(T) reference, which is a significant reduction from 0.52 kcal/mol for PBE0+TS-vdW.
We find that the main effect of the MBD energy over the pairwise TS-vdW approximation is to
destabilize the extended conformations of Ala4, bringing their energies in much better agreement
with the reference CCSD(T) values.
4.3 Supramolecular Systems: The Buckyball Catcher
Supramolecular host–guest systems play an important role for a wide range of applications in
chemistry and biology. The prediction of the stability of host–guest complexes represents a great
challenge for first-principles calculations due to the interplay of a wide variety of covalent and
non-covalent interactions in these systems. Here we assess the performance of the DFT+MBD
65
Figure 5: Performance of PBE0+TS-vdW and PBE0+MBD for the conformational energies of
Ala4. The reference CCSD(T) energies are taken from Ref. [101].
Figure 6: Illustration of the geometry and the anisotropy in the atomic TS+SCS polarizabilities
of the C60@C60H28 complex. The polarizability tensors are visualized as ellipsoids [104].
66
method on the binding of the so-called “buckyball catcher” complex, C60@C60H28, shown in
Figure 6. Since its synthesis [105], the buckyball catcher has become one of the most widely
used benchmark systems for supramolecular chemistry. Recently a reliable binding energy of
26 ± 2 kcal/mol has been determined for the C60@C60H28 complex from large-scale diffusion
Monte Carlo (DMC) calculations [63]. This value is in excellent agreement with an extrapolated
binding energy determined from the experimentally measured binding affinity [29].
All pairwise-corrected dispersion-inclusive DFT calculations significantly overestimate the sta-
bility of the buckyball catcher complex, anywhere from 9 to 17 kcal/mol [63]. The PBE+MBD
method yields a binding energy of 36 kcal/mol, improving the binding by 7 kcal/mol compared
to the PBE+TS-vdW method. The inclusion of exact exchange using the PBE0+MBD method
leads to a negligible change in the binding energy. Therefore, the PBE0+MBD method overesti-
mates the binding by at least 8 kcal/mol compared to the DMC and extrapolated experimental
reference binding energies.
In order to understand the most likely origin of why the binding energy of C60@C60H28 complex
is overestimated by PBE0+MBD, we show the projected polarizability tensors of the full complex
resulting from the TS+SCS calculation in Figure 6. One can clearly see that the polarizability
distribution is highly anisotropic, with an increasing anisotropy close to the linker moiety that
connects the two corannulene molecules of the catcher complex. While the approximation of
isotropic C6 coefficients used in DFT+MBD becomes sufficient as the distance between the atoms
is increased, at shorter interatomic distances the anisotropy plays a non-negligible role [62].
At present, there is no efficient method that can accurately account for the fully anisotropic
dispersion energy at close interatomic distances. This statement applies to the widely employed
interatomic dispersion energy methods, as well as the non-local density functionals (e.g., different
variants of the vdW-DF method [11]). Work is currently in progress to seamlessly include
anisotropy in dispersion energy expressions [62,77]. The anisotropy in the atomic polarizabilities
will change the vdW energy contribution in different directions. In the case of the C60@C60H28
complex, the polarizability of the C60H28 molecule is highly anisotropic as shown in Figure 6. In
the isotropic approximation, the dispersion energy between the C60 molecule and the corannulene
moieties is therefore overestimated, because the polarization is artificially extended towards the
C60 molecule. The fully anisotropic treatment of the dispersion energy is therefore likely to
bring the binding energy closer to the DMC reference value.
4.4 Molecular Crystals
The understanding and prediction of the structure and stability of molecular crystals is of
paramount importance for a variety of applications, including pharmaceuticals, non-linear optics,
and hydrogen storage [106, 107]. The crystal structure prediction blind tests conducted by
the Cambridge Crystallographic Data Centre have shown steady progress toward theoretical
structure prediction for molecular crystals [108]. However, the insufficiency of DFT with pairwise
dispersion corrections for the reliable predictions of molecular crystals is well documented, see
e.g., Refs. [27, 109–111].
To illustrate the role of MBD interactions in the stability of molecular crystals, we have studied a
series of oligoacene crystals from naphthalene to pentacene. We have recently shown that reliable
67
structures of oligoacene crystals (2% accuracy compared to low-temperature X-ray data) can
be obtained with PBE+TS-vdW calculations, while MBD interactions play only a minor role in
determining the geometry of these molecular crystals [112]. However, the MBD energy plays a
more significant role for the lattice energies of oligoacene crystals. Table 4 shows lattice energies
at 0 K for naphthalene (2 benzene rings), anthracene (3 rings), tetracene (4 rings), and pentacene
(5 rings) calculated using the PBE+TS-vdW and PBE+MBD methods, as well as a range of
measured sublimation enthalpies extrapolated to 0 K. We have only taken those experimental
values that are recommended as reliable after critical revision by the authors of Ref. [113], thus
avoiding anomalously small or large sublimation enthalpies. Both naphthalene and anthracene
crystals have been rigorously studied, and their sublimation enthalpies are well known with a
spread of 0.05 and 0.12 eV per molecule, respectively. There are fewer measurements available
for tetracene and pentacene, and for the latter the three available experimental values deviate
by 0.55 eV per molecule.
For naphthalene, anthracene, and tetracene, the PBE+MBD method decreases and improves the
binding by about 0.1 eV (2.3 kcal/mol) per molecule when compared to PBE+TS-vdW. This is
a notable improvement, especially if viewed in the context of intermolecular interactions for the
S22 and S66 databases. We remind the reader that the errors of PBE+TS-vdW and PBE+MBD
for molecular dimers in the S22/S66 databases are well below 0.5 kcal/mol. The much larger dif-
ference between the pairwise PBE+TS-vdW approach and the many-body PBE+MBD method
for molecular crystals can be explained by the presence of significant electrodynamic screening
effects in extended systems, that are virtually absent in small molecules. We refer the reader
to Ref. [112] for a detailed analysis of the importance of electrodynamic screening in molecular
crystals.
The remaining slight overestimation of lattice energies in Table 4 by PBE+MBD compared to
the experimental range can be explained by the fact that the sublimation enthalpy is measured
at finite temperatures, where the crystal unit cell undergoes thermal expansion. When using the
experimental unit cell at 295 K for naphthalene, the PBE+MBD method yields a lattice energy
that is increased by 50 meV per molecule, which places it essentially within the experimental
range reported in Table 4. Finally, we studied the influence of exact exchange for oligoacene
crystals, finding that the PBE0+MBD method leads to an almost negligible difference when
compared to PBE+MBD; the lattice energy is decreased by only 10 meV/molecule when the
PBE0+MBD functional is employed instead of PBE+MBD.
Table 4: Lattice energies of oligoacene crystals including zero-point energy (PBE+TS-vdW and
PBE+MBD calculations were carried out using optimized PBE+TS-vdW geometries). The
range of experimental (“Exp.”) “lattice energies” from Ref. [113] and extrapolated to 0 K. All
values are in units of eV per molecule.
PBE+TS-vdW PBE+MBD Exp.
naphthalene -0.950 -0.862 -0.803 to -0.752
anthracene -1.324 -1.206 -1.148 to -1.024
tetracene -1.666 -1.587 -1.525 to -1.299
pentacene -2.035 -2.018 -2.082 to -1.533
68
Our current work on a broad dataset of molecular crystals and their polymorphs [114,115] shows
that beyond-pairwise many-body vdW interactions can be even more significant than found here
for the oligoacene crystals.
5 Remaining Challenges
In this highlight we have described a recently developed method for the many-body vdW
dispersion energy based on a system of quantum harmonic oscillators (QHO). The resulting
DFT+MBD approach is parameter-free for the determination of the frequency-dependent po-
larizability, and uses a single range-separation parameter for the coupling between the long-range
many-body vdW energy and a given DFT functional. We view the DFT+MBD model as a cru-
cial first step in the development of a reliable (accurate and efficient) method for describing
many-body vdW interactions in complex materials.
Currently, the DFT+MBD method essentially amounts to solving the ACFD-RPA correlation
energy equation for a system of localized screened QHOs in the dipole (long-range) approxi-
mation. There are several important extensions that can be accomplished within the ACFD
framework that would allow us to go beyond the DFT+MBD method:
1. Improving the dipolar response. The TS+SCS method defined in Eq. (16) yields
the full non-local interacting response matrix as a function of atomic positions r and r′.
Currently, this information is not fully utilized in the DFT+MBD approach, since we use
contracted isotropic TS+SCS atomic polarizabilities as input for the ACFD-RPA formula.
In principle, the full response matrix can be used in the ACFD-RPA expression, however
this requires a matching definition for the range-separated Coulomb potential. The in-
teracting TS+SCS response matrix transforms the original atom-based representation to
an eigenvector representation for the coupled modes of the system. The Coulomb inter-
action between the coupled modes needs to be extended from our current definition of
range-separation that is based on atomic vdW radii.
2. Going beyond the dipole approximation. The QHO model possesses a response to
infinite order in the multipole expansion. The current MBD method restricts the response
to the dipole approximation, effectively allowing excitations only to the first excited state
for every QHO due to the dipole selection rule. In principle, the full response function
given by Eq. (2) can be computed for a system of QHOs up to an arbitrary energy cutoff
for the excited states. This would allow us to treat multipole responses higher than dipole
(quadrupole, octupole, etc.). The ACFD-RPA expression can still be utilized in this case,
allowing us to compute dispersion interactions at shorter interatomic distances. It remains
to be assessed whether or not this model will be useful, as a single QHO per atom might not
be able to properly describe vdW interactions at shorter interatomic distances. However,
in principle, our method can also be extended to represent every atom by several QHOs.
3. Coupling between the long-range vdW energy and the DFT energy. The
DFT+MBD method couples the long-range vdW energy to the DFT energy by using
a single range-separation parameter in the Coulomb potential. In order to improve this
69
empirical component of the DFT+MBD method, the DFT functional has to be derived
in the presence of the long-range vdW energy. To date, we have not used the fact that
different functionals yield different results for the electron density tails; this information
can be useful for developing a functional in which the long-range vdW energy is seamlessly
integrated with the semilocal exchange-correlation functional.
4. Simultaneous description of localized and metallic states. Successful non-empirical
DFT functionals are based on the local-density approximation (LDA) and converge to the
LDA in the homogeneous electron gas (HEG) limit. LDA is an exact functional for the
HEG, hence it includes vdW interactions inside the HEG. Therefore, a seamless vdW func-
tional should yield a vanishing correction for the HEG. This can easily be accomplished by
letting the polarizability vanish for slowly-varying regions of the electron density, as done
in the vdW-DF [22] and VV10 [92] approaches. However, real materials (transition metals,
nanostructures, etc.) are more complex than the rather simplified HEG model. In such
systems, vdW interactions between ions are significant and are screened by the itinerant
metallic electrons [116]. State-of-the-art vdW functionals do not correctly describe this
complex situation. However, the DFT+MBD method can be extended to systems with
localized and metallic states by introducing both localized and delocalized oscillators for
every atom. The challenge consists of defining the oscillator parameters directly from the
electron density and its gradient.
5. Interatomic forces, geometry optimization, and molecular dynamics. Currently,
the DFT+MBD method only yields the total energy for a specified geometry. In principle,
geometry optimizations are possible by using the finite difference approximation for the
interatomic forces. This is, however, computationally expensive especially in the case of
molecular dynamics. Work is in progress to derive an analytic expression for the inter-
atomic forces corresponding to the MBD energy. Such development would allow for the
routine application of the DFT+MBD method in large-scale molecular dynamics simula-
tions.
6 Conclusions
There is mounting evidence that many-body vdW interactions, beyond the standard pairwise
approximation, play a crucial role in the structure, stability, and function of a wide variety of sys-
tems of importance in biology, chemistry, and physics. In this highlight, we have illustrated the
importance of including many-body vdW interactions when describing small molecular dimers,
conformational energies of peptides, binding in supramolecular systems, and cohesion in molec-
ular crystals. We presented a derivation of both the pairwise and many-body interatomic vdW
dispersion energy from the exact quantum-mechanical ACFD-RPA correlation energy expres-
sion. The ACFD formula provides us with a powerful framework for the understanding and
future development of accurate and efficient electronic structure approaches.
The DFT+MBD method [32, 33] represents a first step towards the development of reliable
methods for describing many-body vdW interactions in complex materials. In this work, we
derived the MBD energy expression from the exact ACFD formula, discussed the approximations
70
involved, and identified the remaining challenges that need to be addressed in future work. Over
the next few years, we anticipate extensive development of new dispersion energy methods that
will address the truly collective many-body nature of these ubiquitous quantum-mechanical
forces.
7 Acknowledgements
The authors acknowledge the European Research Council (ERC Starting Grant VDW-CMAT)
for support, and thank M. Scheffler, R. Car, X. Ren, J. F. Dobson, O. A. von Lilienfeld, A.
Ambrosetti, A. M. Reilly, and N. Marom for enlightening discussions. R.A.D. was supported by
the DOE under Grant No. DE-SC0005180 and by the NSF under Grant No. CHE-0956500.
References
[1] V. A. Parsegian, Van der Waals forces: A Handbook for Biologists, Chemists, Engineers
and Physicists (Cambridge University Press, 2005).
[2] A. J. Stone, The Theory of Intermolecular Forces (Oxford University Press, 1996).
[3] I. G. Kaplan, Intermolecular Interactions: Physical Picture, Computational Methods and
Model Potentials (John Wiley and Sons, 2006).
[4] R. French et al., Rev. Mod. Phys. 82, 1887 (2010).
[5] B. Jeziorski, R. Moszynski and K. Szalewicz, Chem. Rev. 94, 1887 (1994).
[6] P. Loskill, H. Hahl, T. Faidt, S. Grandthyll, F. Muller and K. Jacobs, Adv. Coll. Interf.
Sci. 107, 179182 (2012).
[7] P. Loskill, J. Puthoff, M. Wilkinson, K. Mecke, K. Jacobs and K. Autumn, J. R. Soc.
Interface (2013), doi:10.1098/rsif.2012.0587.
[8] K. E. Riley, M. Pitonak, P. Jurecka and P. Hobza, Chem. Rev. 110, 5023 (2010).
[9] F. O. Kannemann and A. D. Becke, J. Chem. Theory Comput. 6, 1081 (2010).
[10] S. Grimme, Comput. Mol. Sci. 1, 211 (2011).
[11] V. R. Cooper, L. Kong and D. C. Langreth, Physics Procedia 3, 1417 (2010).
[12] S. Steinmann and C. Corminboeuf, J. Chem. Theory Comput. 7, 3567 (2011).
[13] A. Tkatchenko, L. Romaner, O. T. Hofmann, E. Zojer, C. Ambrosch-Draxl and M. Schef-
fler, MRS Bull. 35, 435 (2010).
[14] A. J. Misquitta, R. Podeszwa, B. Jeziorski and K. Szalewicz, J. Chem. Phys. 123, 214103
(2005).
71
[15] O. A. von Lilienfeld, I. Tavernelli, U. Rothlisberger and D. Sebastiani, Phys. Rev. Lett.
93, 153004 (2004).
[16] E. R. Johnson and A. D. Becke, J. Chem. Phys. 123, 024101 (2005).
[17] P. L. Silvestrelli, Phys. Rev. Lett. 100, 053002 (2008).
[18] G. Mercurio, E. R. McNellis, I. Martin, S. Hagen, F. Leyssner, S. Soubatch, J. Meyer,
M. Wolf, P. Tegeder, F. S. Tautz and K. Reuter, Phys. Rev. Lett. 104, 036102 (2010).
[19] N. Atodiresei, V. Caciuc, P. Lazic and S. Blugel, Phys. Rev. Lett. 102, 136809 (2009).
[20] K. Tonigold and A. Gross, J. Chem. Phys. 132, 224701 (2010).
[21] D. Stradi, S. Barja, C. Diaz, M. Garnica, B. Borca, J. J. Hinarejos, D. Sanchez-Portal,
M. Alcami, A. Arnau, A. L. Vazquez de Parga, R. Miranda and F. Martin, Phys. Rev.
Lett. 106, 186102 (2011).
[22] M. Dion, H. Rydberg, E. Schroder, D. C. Langreth and B. I. Lundqvist, Phys. Rev. Lett.
92, 246401 (2004).
[23] K. Lee, E. D. Murray, L. Kong, B. I. Lundqvist and D. C. Langreth, Phys. Rev. B 92,
081101 (2010).
[24] O. A. Vydrov and T. Van Voorhis, J. Chem. Theory Comput. 8, 1929 (2012).
[25] A. Tkatchenko and M. Scheffler, Phys. Rev. Lett. 102, 073005 (2009).
[26] N. Marom, A. Tkatchenko, M. Rossi, V. V. Gobre, O. Hod, M. Scheffler and L. Kronik,
J. Chem. Theory Comput. 7, 3944 (2011).
[27] O. A. von Lilienfeld and A. Tkatchenko, J. Chem. Phys. 132, 234109 (2010).
[28] S. Grimme, J. Antony, S. Ehrlich and H. Krieg, J. Chem. Phys. 132, 154104 (2010).
[29] S. Grimme, Chem. Eur. J 18, 9955 (2012).
[30] G.-X. Zhang, A. Tkatchenko, J. Paier, H. Appel and M. Scheffler, Phys. Rev. Lett. 107,
245501 (2011).
[31] V. G. Ruiz, W. Liu, E. Zojer, M. Scheffler and A. Tkatchenko, Phys. Rev. Lett. 108,
146103 (2012).
[32] A. Tkatchenko, R. A. DiStasio, Jr., R. Car and M. Scheffler, Phys. Rev. Lett. 108, 236402
(2012).
[33] R. A. DiStasio Jr., O. A. von Lilienfeld and A. Tkatchenko, Proc. Natl. Acad. Sci. USA
109, 14791 (2012).
[34] J. F. Dobson and T. Gould, Calculation of Dispersion Energies,
http://www.psi-k.org/newsletters/News_107/Highlight_107.pdf, 2011.
[35] J. Klimes and A. Michaelides, J. Chem. Phys. 137, 120901 (2012).
72
[36] Towards First-Principles Description of van der Waals Interactions in Complex Materials,
http://www.cecam.org/workshop-791.html, 2012.
[37] O. Gunnarsson and B. I. Lundqvist, Phys. Rev. B 13, 4274 (1976).
[38] D. C. Langreth and J. P. Perdew, Phys. Rev. B 15, 2884 (1977).
[39] S. L. Adler, Phys. Rev. 126, 413 (1962).
[40] N. Wiser, Phys. Rev. 129, 62 (1963).
[41] D. Bohm and D. Pines, Phys. Rev. 92, 609 (1953).
[42] M. Gell-Mann and K. A. Brueckner, Phys. Rev. 106, 364 (1957).
[43] M. Fuchs and X. Gonze, Phys. Rev. B 65, 235109 (2002).
[44] F. Furche and T. van Voorhis, J. Chem. Phys. 122, 164106 (2005).
[45] F. Furche, J. Chem. Phys. 129, 114105 (2008).
[46] G. E. Scuseria, T. M. Henderson and D. C. Sorensen, J. Chem. Phys. 129, 231101 (2008).
[47] B. G. Janesko, T. M. Henderson and G. E. Scuseria, J. Chem. Phys. 130, 081105 (2009).
[48] J. Toulouse, I. C. Gerber, G. Jansen, A. Savin and J. G. Angyan, Phys. Rev. Lett. 102,
096404 (2009).
[49] J. Harl and G. Kresse, Phys. Rev. B 77, 045136 (2008).
[50] J. Harl and G. Kresse, Phys. Rev. Lett. 103, 056401 (2009).
[51] D. Lu, Y. Li, D. Rocca and G. Galli, Phys. Rev. Lett. 102, 206411 (2009).
[52] J. F. Dobson and J. Wang, Phys. Rev. Lett 82, 2123 (1999).
[53] M. Rohlfing and T. Bredow, Phys. Rev. Lett 101, 266106 (2008).
[54] X. Ren, P. Rinke and M. Scheffler, Phys. Rev. B 80, 045402 (2009).
[55] L. Schimka, J. Harl, A. Stroppa, A. Gruneis, M. Marsman, F. Mittendorfer and G. Kresse,
Nature Mater. 9, 741 (2010).
[56] Y. Li, D. Lu, H.-V. Nguyen and G. Galli, J. Phys. Chem. A 114, 1944 (2010).
[57] X. Ren, A. Tkatchenko, P. Rinke and M. Scheffler, Phys. Rev. Lett. 106, 153003 (2011).
[58] X. Ren, P. Rinke, C. Joas and M. Scheffler, J Mater. Sci. 47, 7447 (2012).
[59] D. Lu, H.-V. Nguyen and G. Galli, J. Chem. Phys. 133, 154110 (2010).
[60] X. Ren, P. Rinke, V. Blum, J. Wieferink, A. Tkatchenko, A. Sanfilippo, K. Reuter and
M. Scheffler, New. J. Phys 14, 053020 (2012).
73
[61] J. F. Dobson, Quasi-local-density approximation for a van der Waals energy functional,
in Topics in Condensed Matter Physics, edited by M. P. Das, p. 121, Nova (New York),
1994.
[62] A. Krishtal, K. Vannomeslaeghe, D. Geldof, C. V. Alsenoy and P. Geerlings, Phys. Rev.
A 83, 024501 (2011).
[63] A. Tkatchenko, D. Alfe and K. S. Kim, J. Chem. Theory Comput. 8, 4317 (2012).
[64] B. M. Axilrod and E. Teller, J. Chem. Phys. 11, 299 (1943).
[65] T. Helgaker, P. Jørgensen and J. Olsen, Molecular Electronic Structure Theory (Wiley,
Chichester, 2000).
[66] A. Mayer, Phys. Rev. B 75, 045407 (2007).
[67] D. W. Oxtoby and W. M. Gelbart, Mol. Phys. 29, 1569 (1975).
[68] B. T. Thole, Chem. Phys. 59, 341 (1981).
[69] R. G. Parr and W. Yang, Density Functional Theory of Atoms and Molecules (Oxford
Science Publications, 1989).
[70] W. Koch and M. C. Holthausen, A Chemist’s Guide to Density Functional Theory (Wiley-
VCH, 2002).
[71] F. L. Hirshfeld, Theor. Chim. Acta 44, 129 (1977).
[72] W. L. Bade, J. Chem. Phys. 27, 1280 (1957).
[73] J. Cao and B. J. Berne, J. Chem. Phys. 97, 8628 (1992).
[74] A. G. Donchev, J. Chem. Phys. 125, 074713 (2006).
[75] M. W. Cole, D. Velegol, H.-Y. Kim and A. A. Lucas, Mol. Simul. 35, 849 (2009).
[76] H. Berthoumieux and A. C. Maggs, Europhys. Lett. 91, 56006 (2010).
[77] A. Tkatchenko, A. Ambrosetti and R. A. DiStasio, Jr., arXiv:1210.8343 (2012).
[78] T. W. Whitfield and G. J. Martyna, Chem. Phys. Lett. 424, 409 (2006).
[79] A. Jones, A. Thompson, J. Crain, M. H. Muser and G. J. Martyna, Phys. Rev. B 79,
144119 (2009).
[80] P. Jurecka, J. Sponer, J. Cerny and P. Hobza, Phys. Chem. Chem. Phys. 8, 1985 (2006).
[81] T. Takatani, E. G. Hohenstein, M. Malagoli, M. S. Marshall and C. D. Sherrill, J. Chem.
Phys. 132, 144104 (2010).
[82] J. P. Perdew, K. Burke and M. Ernzerhof, Phys. Rev. Lett. 77, 3865 (1996).
[83] J. P. Perdew, M. Ernzerhof and K. Burke, J. Chem. Phys. 105, 9982 (1996).
74
[84] C. Adamo and V. Barone, J. Chem. Phys. 110, 6158 (1999).
[85] V. V. Gobre et al., in preparation.
[86] R.-F. Liu, J. G. Angyan and J. F. Dobson, J. Chem. Phys. 134, 114106 (2011).
[87] J. R. Hammond, M. Valiev, W. A. de Jong and K. Kowalski, J. Phys. Chem. A 111, 5492
(2007).
[88] J. R. Hammond, K. Kowalski and W. A. de Jong, J. Chem. Phys. 127, 144105 (2007).
[89] S. Botti, A. Castro, X. Andrade, A. Rubio and M. A. L. Marques, Phys. Rev. B 78,
035333 (2008).
[90] A. Kumar, M. Kumar and W. J. Meath, Chem. Phys. 286, 227 (2003).
[91] V. Blum, R. Gehrke, F. Hanke, P. Havu, V. Havu, X. Ren, K. Reuter and M. Scheffler,
Comp. Phys. Commun. 180, 2175 (2009).
[92] O. A. Vydrov and T. Van Voorhis, J. Chem. Phys. 133, 244103 (2010).
[93] J. Rezac, K. E. Riley and P. Hobza, J. Chem. Theory Comput. 7, 2427 (2011).
[94] W. Chin, F. Piuzzi, I. Dimicoli and M. Mons, Phys. Chem. Chem. Phys. 8, 1033 (2006).
[95] J. P. Simons, Mol. Phys. 107, 2435 (2009).
[96] F. Bierau, P. Kupser, G. Meijer and G. von Helden, Phys. Rev. Lett. 105, 133402 (2010).
[97] M. P. Gaigeot, Phys. Chem. Chem. Phys. 12, 3336 (2010).
[98] J. A. Stearns, C. Seaiby, O. V. Boyarkin and T. R. Rizzo, Phys. Chem. Chem. Phys. 11,
125 (2009).
[99] M. Rossi, V. Blum, P. Kupser, G. von Helden, F. Bierau, K. Pagel, G. Meijer and M. Schef-
fler, J. Phys. Chem. Lett. 1, 3465 (2010).
[100] R. L. Baldwin, J. Mol. Biol. 371, 283 (2007).
[101] A. Tkatchenko, M. Rossi, V. Blum, J. Ireta and M. Scheffler, Phys. Rev. Lett. 106, 118102
(2011).
[102] R. A. DiStasio Jr., Y. Jung and M. Head-Gordon, J. Chem. Theory Comput. 1, 862
(2005).
[103] R. A. DiStasio Jr., R. P. Steele, Y. M. Rhee, Y. Shao and M. Head-Gordon, J. Comp.
Chem. 28, 839 (2007).
[104] K. N. Trueblood, H.-B. Burgi, H. Burzlaff, J. D. Dunitz, C. M. Gramaccioli, H. H. Schulz,
U. Shmueli and S. C. Abrahams, Acta Crys. A 52, 770 (1996).
[105] A. Sygula, F. R. Fronczek, R. Sygula, P. W. Rabideau and M. M. Olmstead, J. Am.
Chem. Soc. 129, 3842 (2007).
75
[106] J. Bernstein, Polymorphism in Molecular Crystals (Oxford University Press, USA (New
York), 2002).
[107] S. L. Price, Phys. Chem. Chem. Phys. 10, 1996 (2008).
[108] D. A. Bardwell et al., Acta. Cryst. B 67, 535 (2011).
[109] G. J. O. Beran and K. Nanda, J. Phys. Chem. Lett. 1, 3480 (2010).
[110] S. Wen, K. Nanda, Y. Huang and G. Beran, Phys. Chem. Chem. Phys. 14, 7578 (2012).
[111] K. Hongo, M. A. Watson, R. S. Sanchez-Carrera, T. Iitaka and A. Aspuru-Guzik, J. Phys.
Chem. Lett. 1, 1789 (2010).
[112] B. Schatschneider, J.-J. Liang, A. M. Reilly, N. Marom, G.-X. Zhang and A. Tkatchenko,
arXiv:1211.1683.
[113] M. V. Roux, M. Temprado, J. S. Chickos and Y. Nagano, J. Phys. Chem. Ref. Data 37,
1855 (2008).
[114] A. M. Reilly and A. Tkatchenko, in preparation.
[115] N. Marom, R. A. DiStasio, Jr., V. Atalla, S. Levchenko, J. R. Chelikowsky, L. Leiserowitz
and A. Tkatchenko, arXiv:1210.5636.
[116] J. J. Rehr, E. Zaremba and W. Kohn, Phys. Rev. B 12, 2062 (1975).
76